This project is a part of S2ORC. For S2ORC, we convert PDFs to JSON using Grobid and a custom TEI.XML to JSON parser. That TEI.XML to JSON parser (grobid2json) is made available here. We additionally process LaTeX dumps from arXiv. That parser (tex2json) is also made available here.

The S2ORC github page includes a JSON schema, but it may be easier to understand that schema based on the python classes in doc2json/s2orc.py.

This custom JSON schema is also used for the CORD-19 project, so those who have interacted with CORD-19 may find this format familiar.

Possible future components (no promises):

• Linking bibliography entries (bibliography consolidation) to papers in S2ORC

NOTE: Conda is shown but any other python env manager should be fine

Go here to install the latest version of miniconda.

Then, create an environment:

conda create -n doc2json python=3.8 pytest
conda activate doc2json
pip install -r requirements.txt
python setup.py develop

The current grobid2json tool uses Grobid to first process each PDF into XML, then extracts paper components from the XML.

Build from the latest Grobid repo or use Grobid docker if you have problems on installing Gorbid.

You will need to have Java (tested with Java 11) installed on your machine. Then, you can install your own version of Grobid and get it running, or you can run the following script:

bash scripts/setup_grobid.sh

Note: before running this script, make sure the paths match your installation path. Else it will fail to install.

This will setup Grobid, currently hard-coded as version 0.7.3. Then run:

bash scripts/run_grobid.sh

to start the Grobid server. Don't worry if it gets stuck at 87%; this is normal and means Grobid is ready to process PDFs.

The expected port for the Grobid service is 8070, but you can change this as well. Make sure to edit the port in both the Grobid config file as well as grobid/grobid_client.py.

There are a couple of test PDFs in tests/input/ if you'd like to try with that.

For example, you can try:

python doc2json/grobid2json/process_pdf.py -i tests/pdf/N18-3011.pdf -t temp_dir/ -o output_dir/

This will generate a JSON file in the specified output_dir. If unspecified, the file will be in the output/ directory from your path.

If you want to process LaTeX, in addition to installing Grobid, you also need to install the following libraries:

• latexpand (apt install texlive-extra-utils)
• tralics (apt install tralics)

To process LaTeX, all files must be in a zip file, similar to the *.gz files you can download from arXiv.

Like PDF, first start Grobid using the run_grobid.sh script. Then, try to process one of the test files available under tests/latex/. For example, you can try:

python doc2json/tex2json/process_tex.py -i test/latex/1911.02782.gz -t temp_dir/ -o output_dir/

Again, this will produce a JSON file in the specified output_dir.

Why do you need Grobid? We use the Grobid citation and author APIs to convert raw strings into structured forms.

To process JATS XML, try:

python doc2json/jats2json/process_jats.py -i test/jats/PMC5828200.nxml -o output_dir/

This will create a JSON file with the same paper id in the specified output directory.

The format of S2ORC releases have drifted over time. Use the load_s2orc function in doc2json/s2orc.py to try and load historic and currect S2ORC JSON.

To process PDFs, you will first need to start Grobid (defaults to port 8070). If you are processing LaTeX, no need for this step.

bash scripts/run_grobid.sh

Then, start the Flask app (defaults to port 8080).

python doc2json/flask/app.py

Go to localhost:8080 to upload and process papers.

Or alternatively, you can do things like:

curl localhost:8080/ -F file=@tests/pdf/N18-3011.pdf

If you use this utility in your research, please cite:

@inproceedings{lo-wang-2020-s2orc,
    title = "{S}2{ORC}: The Semantic Scholar Open Research Corpus",
    author = "Lo, Kyle  and Wang, Lucy Lu  and Neumann, Mark  and Kinney, Rodney  and Weld, Daniel",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.447",
    doi = "10.18653/v1/2020.acl-main.447",
    pages = "4969--4983"
}

Contributions are welcome. Note the embarassingly poor test coverage. Also, please note this pipeline is not perfect. It will miss text or make errors on most PDFs. The current PDF to JSON step uses Grobid; we may replace this with a different model in the future.

Issues: contact [email protected] or [email protected]